Learning Like Humans: Advancing LLM Reasoning Capabilities via Adaptive Difficulty Curriculum Learning and Expert-Guided Self-Reformulation
This work addresses the problem of inconsistent performance on complex reasoning tasks for LLM users, representing a strong domain-specific advancement rather than a foundational breakthrough.
The paper tackles the challenge of improving large language models' ability to solve complex mathematical reasoning problems by proposing two human-inspired strategies: Adaptive Difficulty Curriculum Learning (ADCL) to dynamically adjust problem difficulty during training, and Expert-Guided Self-Reformulation (EGSR) to guide models in reformulating expert solutions. The combined approach achieved performance improvements of 10% on AIME24 and 16.6% on AIME25 benchmarks over a baseline.
Despite impressive progress in areas like mathematical reasoning, large language models still face significant challenges in consistently solving complex problems. Drawing inspiration from key human learning strategies, we propose two novel strategies to enhance the capability of large language models to solve these complex problems. First, Adaptive Difficulty Curriculum Learning (ADCL) is a novel curriculum learning strategy that tackles the Difficulty Shift phenomenon (i.e., a model's perception of problem difficulty dynamically changes during training) by periodically re-estimating difficulty within upcoming data batches to maintain alignment with the model's evolving capabilities. Second, Expert-Guided Self-Reformulation (EGSR) is a novel reinforcement learning strategy that bridges the gap between imitation learning and pure exploration by guiding models to reformulate expert solutions within their own conceptual framework, rather than relying on direct imitation, fostering deeper understanding and knowledge assimilation. Extensive experiments on challenging mathematical reasoning benchmarks, using Qwen2.5-7B as the base model, demonstrate that these human-inspired strategies synergistically and significantly enhance performance. Notably, their combined application improves performance over the standard Zero-RL baseline by 10% on the AIME24 benchmark and 16.6% on AIME25.